Ubuntu powers the majority of the world's AI workloads today. As artificial intelligence moves from big data centers to our personal laptops, many users feel nervous. They worry about privacy, system bloat, and losing control over their machines.
To address this, Canonical developers are now planing to make AI development safe and fast without forcing it on people who don't want it. They are taking a "principled and focused" approach that prioritises local privacy and user choice.
Table of Contents
The Two Categories of AI on Ubuntu
According to Jon Seager, VP of Engineering at Canonical, Ubuntu divides AI into two categories to keep the experience predictable:
1. Implicit AI
These tools work in the background to improve existing features. A key example is bringing high-quality speech-to-text and text-to-speech to Ubuntu to improve accessibility. You won't see a "bot"; the OS just becomes more capable.
2. Explicit AI
These are new, optional workflows. They include "agentic" tools that can help you author documents, automate daily news briefings, or even troubleshoot a broken Wi-Fi connection.
Starting with Ubuntu 26.10, these features will be a "preview" and strictly opt-in.
No More Driver Struggles
In the past, setting up a Linux machine for AI was a chore. You had to search for drivers and add risky third-party repositories.
The upcoming Ubuntu versions fix this by moving specialized toolkits directly into the official archive. So, you can simply run apt install cuda or apt install rocm on a base system.
Since these drivers are part of the official OS, Canonical provides security updates and guarantees they will work with your hardware for the long haul. In fact, for enterprise users, Canonical offers security maintenance for 15 years on these stacks.
Inference Snaps: Models That Fit Your PC
Finding and downloading right AI models for your hardware might be a challenging task. Inference Snaps automate this.
When you install a snap for a model like DeepSeek-R1 or Qwen, the system detects your hardware and automatically fetches the version optimised by the silicon manufacturer (like NVIDIA, AMD, or Intel) for your specific chip.
Every inference snap includes a local API that follows the OpenAI standard. This means you can immediately plug your local model into developer tools like the VS Code "Continue" extension.
Most importantly, these snaps use strict confinement. They run in a digital sandbox that blocks the model from looking at your private data without your permission.
Protecting Your Ubuntu System with Sandboxing
The "Agentic" AI tools can run commands and modify files to finish tasks. This is powerful, but it's risky. An agent could accidentally delete your home directory.
Ubuntu provides three ways to limit this "blast radius":
- LXD Containers: Run an agent in a disposable environment. You can mount your project folder into the container, let the agent work, and if it makes a mistake, your main OS is safe.
- Virtual Machines: For more isolation, use LXD VMs or Multipass. These provide a separate kernel between the AI and your machine.
- Audit Trails: Ubuntu uses existing system "primitives" like strict access controls and audit trails to track every decision an AI agent makes.
Privacy and the "Killswitch"
Canonical’s policy is local inference by default. Your logs and data stay on your machine. If you want to use a cloud service, you must manually provide your own API token.
While there isn't a single "global killswitch", Ubuntu gives you total control through its packaging. Since these AI features are delivered as Snaps, you can remove them individually at any time. If you don't install the snaps, the AI features simply aren't there.
Starting with Ubuntu 26.10, the AI features will be a strictly opt-in preview. You can even choose to skip them during the initial setup wizard.
What do you think of the Canonical's decision to include AI/LLM features in Ubuntu? Share your opinions in the comment section below.
Related Read:
